Ep. 041 - EMERGENCY: Inside the Anthropic’s S-1 (Tokenomics/ Markets)

Summary

• Anthropic’s disclosed economics challenge the idea that frontier inference depends on subsidies. Joey cites two profitable quarters, margins above 10%, 85% inference gross margins and 41% compute margins. Jeremie highlights 90% gross margins and a potential 40% EBIT profile, retaining the condition “if they get there”: “You’re selling a dollar for 10 cents.”
• The compute commitments look manageable to the panel because liquidity and revenue have grown alongside them. Jeremie contrasts $500–600 billion of commitments with over $100 billion already on the books and prospective funding taking resources above $200 billion. His reassurance—“they’re not going to go under”—does not settle whether revenue will accelerate enough to monetize the 2027 capacity.
• Customer expansion supports the growth story, but concentration still matters. Joey describes existing-customer spending rising ninefold, versus fivefold in Q1; the discussion ultimately separates the top five customers’ 25% revenue contribution from 48% through cloud marketplaces. Jordan calls purchasing commitments and slow model migration “really sticky”: 20% of Q3 revenue still came from models released before April.
• SpaceX’s expensive, flexible compute can benefit from financing pressure on competing neoclouds. Doug describes a “liquidity run” on neocloud debt and interprets Musk’s criticism as competitive positioning. Jeremie says $14/hour GB300 pricing could persist or strengthen if demand outgrows even substantial supply additions; short commitments also conserve Anthropic’s balance-sheet capacity despite lower margins.
• Compute efficiency strengthens Anthropic’s economics, although the panel disputes how much credit belongs to the company. Joey cites inference revenue per megawatt rising from $16 million to over $60 million. Doug asks how much reflects Anthropic versus broader hardware gains; Jordan and Doug emphasize software, accelerator diversity and model-assisted engineering: “the model is the differentiator.”
• Enterprise budgets offer substantial runway, with Q4 seasonality an unresolved counterweight. Joey puts Anthropic at roughly 0.1% of Fortune 500 operating budgets and expects a strong 2027 as formal AI budgeting expands. But the prior holiday period brought a 20% spending decline over two weeks, and Doug’s market assessment is blunt: “The AI stocks are tired.”
• The disagreement on OpenAI concerns competitive economics more than survival. Jordan frames OpenAI as caught between rivals; Jeremie argues weaker “best-in-breed” positioning would force price competition, but says he remains bullish on both and finds OpenAI more reliable for some data analysis. A panelist raises OpenAI’s enterprise sales as a countercase, while another notes its improving execution and the burden of supporting a large free consumer base.

Deep dive

1. Inference margins undermine the subsidy argument

• Joey opens with two profitable quarters, margins above 10%, 85% inference gross margins and 41% compute margins. His explanation combines inference efficiencies with efficient sales, marketing and partnerships: customers expand usage without requiring an equally large acquisition effort. More revenue can therefore flow through without recreating the selling costs associated with winning the original account.

• Jeremie describes the incremental economics as 90% gross margins and a potential 40% EBIT profile, “if they get there.” His rebuttal to the subsidy thesis is deliberately simple: “You’re selling a dollar for 10 cents.” The significance is that customers genuinely want the output and pay substantially more than its delivery cost.

• Joey’s retention example supplies the growth mechanism: a customer spending $1 a year earlier now spends $9; in Q1, the comparable figure was $5. That expansion requires little additional selling. “There’s not much you have to do,” he says of upselling usage, connecting customer behavior directly to the margin story.

2. Large commitments look different against a funded balance sheet

• Jeremie’s data-center perspective starts with creditworthiness. Developers have wanted to sell Anthropic capacity but could not readily finance projects against its credit; the IPO opportunity remains contingent on investment-grade standing. He cites over $100 billion on the books and expects IPO proceeds and promised funding to take resources above $200 billion, after a period when the company “barely burned any cash.”

• The $500–600 billion compute headline needs a denominator, Jeremie argues. At roughly $100 billion per gigawatt in the Nscale comparison, half a trillion dollars would purchase only 5 GW. Anthropic has contracted materially more capacity than that, implying a comparatively low commitment burden per gigawatt. Its current footprint is under 5 GW, with a substantial increase expected by the end of next year.

• His comparison with the earlier OpenAI scare is roughly $80 billion of revenue against half a trillion dollars of commitments, rather than $20 billion annualized against $1.4 trillion. Jeremie forecasts Anthropic reaching “multiple hundreds of billions” in annual revenue next year. He also emphasizes the mix of long contracts and flexible arrangements with hyperscalers and SpaceX, rather than treating every commitment as equally binding.

• Further funding includes Google’s $40 billion package, with another $20 billion tied to milestones such as going public, and a $42 billion Broadcom arrangement that Jeremie says enables more TPU purchases. He regards much of the 2027 compute roadmap as de-risked. The remaining question is explicit: “Can they actually accelerate the revenue enough?” Funding capacity establishes resilience; monetization still has to follow.

3. Customer concentration and distribution require separate readings

• Jeremie initially wonders what remains if the top-five and cloud-marketplace revenue shares add together. Jordan explains that the categories can overlap, and Doug probes whether the customer figures are look-through. Joey clarifies that the 48% of revenue from CSP marketplaces excludes the top five customers. The discussion settles on 25% from the top five, 48% through cloud marketplaces and 27% elsewhere.

• Jordan cites alternative-data panels in which the top 1% of customers account for 70–80% of the panel’s spending. That is panel evidence, not Anthropic’s disclosed customer mix. Jordan also cites 250,000 customers on a trailing-12-month basis and says SemiAnalysis is one of 1,900 in the “million-plus” group. He and Joey note that the top 5% of Claude Code users spend $18,000 annually, while non-Claude Code users spend under $75 monthly.

• Jordan’s stickiness argument combines purchasing commitments with slow migration: 20% of Q3 revenue came from models released before April. Fortune 500 buyers using Bedrock may need certifications before upgrading, whereas startups can move faster. He treats that lag and AWS-style purchasing agreements as evidence that relationships are “really sticky.” Speculation about the identities of anonymous major customers remains speculation, with several suggestions disputed rather than resolved.

4. Neocloud financing pressure can sustain SpaceX’s premium

• Doug interprets Musk’s attacks on neoclouds as competitive behavior from someone operating a neocloud himself. A “liquidity run on all the neocloud debt” impairs rivals’ ability to finance capacity and can redirect investors toward SpaceX, lowering its cost of capital. A panelist warns that data centers are more constrained than chips today, but that could reverse by the end of next year as Musk pursues 10–20 GW.

• Asked whether $14/hour GB300 contracts cancellable monthly can persist, Jeremie makes the answer conditional on demand. Supply growth should be huge, but if revenue and compute demand outperform it, premiums could increase: “Who else has as much capacity available in 2027?” If demand fades, the economics become harder. Financing constraints on smaller suppliers make an unlimited competitive supply response difficult even when attractive prices exist.

• Contract duration explains why Anthropic might accept those prices. Jeremie compares six-year commitments around $100 billion per gigawatt, potentially rising to $120 billion or more, with SpaceX pricing around $50 million per megawatt annually but a 90-day commitment—roughly $15 billion per gigawatt committed. “Margins go down,” he acknowledges, while the recorded obligation stays much smaller. Jeremie also notes the risks of equity-funding capacity and building without a customer.

5. Better hardware and better software reinforce each other

• Joey cites inference revenue per megawatt rising from about $16 million last year to over $60 million currently. The Q3 dip reflects new SpaceX capacity, less Meta distillation revenue and additional safety compute. Doug reads another S-1 comparison: revenue grew 1,368% year over year for the three months ended June 30, while compute and infrastructure expenses grew 300%. Their conclusion is that inference scales unusually well.

• Doug asks how much of the efficiency reflects Anthropic versus broader hardware gains, including a cited 50× GB-versus-Hopper comparison. Jordan answers that newer silicon needs software capable of exploiting it. Anthropic can select among TPUs, Trainium, Nvidia GPUs and AMD GPUs, matching accelerators to workloads rather than depending on one platform. Hardware improvement and software execution therefore work together.

• Jordan says the team may have “underestimated the models,” particularly their ability to perform difficult engineering work. His example is reverse-engineering game binaries that had supposedly been exceptionally resistant to that process. He thinks Fable 5.5 can probably port things across accelerators, reducing the importance of hardware-specific barriers. Diversifying silicon and suppliers also limits individual vendors’ bargaining power: “the model is the differentiator.”

• Jordan sees favorable timing: earlier entrants accumulated older H100 fleets, while new labs lack Anthropic’s revenue, compute, research and data flywheels. “I don’t know how anybody catches up from here.” A panelist flags a relative weakness in multimodal vision compared with Astra; the panel also allows that specialized applications could capture niches. The broader software-and-research advantage does not mean every domain is already won.

6. Enterprise expansion faces a holiday test and a tired market

• Joey sees runway even without turning everyone into an extreme user. The top 5% of Claude enterprise users spend $18,000 annually, while non-Claude Code users spend under $75 monthly. Anthropic represents about 0.1% of Fortune 500 operating budgets. Moving the median employee toward today’s 75th percentile would matter without requiring 90th- or 99th-percentile consumption. “We’re painting the lines,” he says of how early adoption remains.

• The next catalyst is the first Q4 enterprise IT budgeting cycle seriously incorporating AI, despite the two frontier labs already generating roughly $120 billion in combined ARR. Joey expects a strong 2027 and leaves 2028 for later assessment. Government contributes approximately 1% of Anthropic revenue. Hundreds of millions of Microsoft 365 seats illustrate the broader opportunity, although ordinary office workers will not immediately spend like coders.

• Jordan flags seasonality: spending fell approximately 20% in the two weeks ended January 2, 2026, versus the two weeks ended December 19—roughly a 10% December impact in the discussion’s shorthand. One panelist expects daily Claude Code habits to soften the next holiday dip; another thinks a shift toward consumption billing could worsen it when users log off. The disagreement remains unresolved rather than becoming a confident Q4 adjustment.

• Doug finds the market response muted relative to the filing’s strength. Data-center land and buildout companies led, but semiconductor and broader AI beneficiaries did not rally as forcefully as earlier in the year: “The AI trade is exhausted.” Doug also highlights management’s statements about not overextending and not missing expectations; a panelist and Jordan frame that posture as “underpromise and overdeliver.” Jeremie, separately, says Q4 numbers being discussed are “stupidly low” and “not believable.”

7. OpenAI’s position draws a genuine disagreement

• Jordan asks whether OpenAI is “cooked” and frames it as caught between Gemini and Meta in consumer, and Anthropic and open models in enterprise. He says Opus 5.5’s gains in SemiAnalysis’s spending appear to come at Astra’s expense. Jeremie argues that a leader can charge a premium, while a weaker “best-in-breed” position leaves OpenAI “competing with Anthropic on price.”

• Jeremie says he remains bullish on both companies and still uses OpenAI models for tasks such as data analysis. Other panelists note that users’ preferences can shift when one model seems better or worse. The disagreement concerns relative positioning and profitability, not whether OpenAI can make a business.

• An unidentified panelist offers a countercase: OpenAI may currently be better at selling to traditional Fortune 500 enterprises. Another panelist points to a strong Q3, greater enterprise focus since July, expected marketplace expansion through Bedrock and Vertex, and aggressive discounting against Anthropic accounts. One- and three-year contracts could reduce rapid model switching. Better sales execution could therefore matter as spending broadens beyond today’s power users.

• A panelist also identifies the consumer burden. If only 6% of a billion weekly active users become paying customers, supporting the remainder is costly. OpenAI has had little usage-based API revenue, which the panel associates with monetization above $100 million per megawatt. A panelist says OpenAI’s margins currently look very different from Anthropic’s, but expects them to improve.